To add a little detail, I'm looking to draw about 750W from a 1500W power supply. This should put it near the peak of its efficiency curve and near the knee of its noise curve.

There are a number of 1500W, 80+ gold (and even platinum) power supplies out there, but I must be looking in all the wrong places for sound pressure level readings. So the question remains...

Which 1500W power supply is the quietest?

Welcome to SPCR. Yep, not many sites providing SPL readings for PSUs reviews above 1kW. I'm a skeptic...is this real 750W draw or a PSU calculator result? Seems like the only way to get that high is with 3-way SLI or crossfire and an overvolted CPU.

Thanks. You're right about that; it seems that most people who want quiet operation are willing to sacrifice performance to get it, and most people who want peak performance don't care how loud the machine gets. As for me, I would like to see how quiet I can make a peak performing machine.

It is an estimated draw for a Core i7 and GTX 580, arrived at as follows:

mobo: 100 W x 1 = 100 W
cpu: 130 W x 1 = 130 W
gpu: 300 W x 1 = 300 W
ram: 0.5 W x 16 = 8 W
disk: 30 W x 1 = 30 W
ssd: 20 W x 1 = 20 W
dvd: 30 W x 1 = 30 W
fan: 3 W x 6 = 18 W
pump: 50 W x 1 = 50 W
overall total: 686 W

But because the power draws are all estimates, I decided to round it to 700W, give it a little head room and call it "about 750W." I had assumed that replacing the 580 with a 680 wouldn't change power budget much, but on looking it up, I see that the 680 actually draws about 100W less than the 580. Nice.

I don't see how that system you linked to gets by with such low power dissipation. The three GTX 680's alone should draw around 600W.

Anyway, the 1500W supplies I've found are the following:

Cooler Master Silent Pro M2 1500W
Enermax MaxREVO 1500W
EVGA SuperNOVA NEX1500 Classified
SilverStone Strider ST1500
Thermaltake Toughpower TP-1500M

Of these, I could only find SPL figures for the EVGA, which are as follows:

150 W: 31.6 dBA
300 W: 32.3 dBA
750 W: 32.8 dBA
1500 W: 64.3 dBA

So, if anyone can add to the list, contribute numbers, or describe experience with with any of these power supply units, I would be most appreciative!

I'd second CA_Steve on this. Just how do you intend to draw 750W? If you really were drawing that kind of power then silence would be irrelevant as the noise produced cooling the rest of the system would drain out all but petrol powered solutions.


This is somewhat of a myth I am afraid. PSUs don't give peak efficiency at 50% power and modern units tend to be best over a broad range, only really dropping off significantly when you're over 90% load. If you wanted to run a system at full power all of the time then the most efficient level would be much higher than 50% and that's if you really do run at 100% usage all of the time. Do you use Prime95 + Furmark all day every day? No. You probably run at very low power levels most of the time so it would be more efficient to buy the lowest rated PSU you need.

Maybe this somewhat goes against what is purported to be the requirements from PSU makers who, let's face it, have a vested interest in you buying a much larger PSU than you need.

It is far more likely that a sub 600W PSU fits your needs.

Thanks for the spec list. Let's get you up to speed on practical power use. Some of your power assumptions are well beyond the norm.



You don't say which core i7 you have...and the only i7's that have 130W TDP are the SNB-E's (or the earlier gen i7-9xx).

My generic response for the rest is:
mobo and RAM ~30W.
GPU TDP 244W
drives + fans call it 20W
water pump ~15W

Total stressed load power ...call it 450W. Total power while gaming <400W. If you overclock/overvolt the CPU and GPU, there's more to discuss. But, you are far far away from drawing 700W or more.

That said, I'd aim for a PSU that's quieter than the rest of the components while under a heavy load. Take a look at the Seasonic x-660, Enermax Platimax 700, bequiet! Dark Power Pro 650, Kingwin LZP-650. All of these models have been reviewed here (at the 560W, 600W, 550W, and 550W variants, respectively) and showed decent performance at 400W load. Getting the next higher wattage model should give you some headroom for the fan profile at 400W. Hope that came across clearly

Sorry, I was referring to the 3960X. According to Intel Core i7-3960X Review: Sandy Bridge-E And X79 Express, the 3960X draws 87W at idle, and according to Core i7-3970X Extreme Review: Can It Stomp An Eight-Core Xeon?, the 3960X draws 160W under full load. These values are averaged over several minutes. The actual peak power draw is close to 400W when averaged over 2-second intervals.

I readily accept your updated power draw for the GTX 680, but I think your motherboard estimates are rather light. Here are some reviews, starting with high-end boards and working down, with idle and loaded power as given by each one:

Overclocking: Asus Rampage IV Extreme Versus EVGA X79 FTW (also gives baseline numbers)
idle: 109-154W
load: 219-265W

Ultimate X79? Five $320+ LGA 2011 Motherboards, Reviewed
idle: 100-115W
load: 225-246W

Seven $260-$320 X79 Express Motherboards, Reviewed
idle: 84-101W
load: 215-241W

Six $200-$260 LGA 2011 Motherboards, Reviewed
idle: 81-95W
load: 206-232W

This makes the last two motherboard reviews appear to violate the laws of physics, because they show some entire boards drawing less than the processor alone at idle. Of course, what this actually shows is that such measurements have a larger margin of error than one might think. In any case, I will round the processor’s power draw down to 80W at idle for consistency.

I’m not interested in the low-end boards, so taking the numbers from the two higher-end reviews, we get
idle: 100-150W (total) – 80W (CPU) = 20-70W (mobo+RAM)
load: 220-250W (total) – 160W (CPU) = 60-90W (mobo+RAM)

Given that, I think budgeting 100W for the motherboard isn’t too extravagant at all.

I don’t see why you don’t like the power estimates for the disks – these numbers were taken directly from the manufacturers’ spec sheets.

The H80/H100 you linked to is a CPU cooler only. I’m considering combined CPU and GPU cooling using a Magicool Triple 180 radiator. That way I can keep the fan and pump speeds to a minimum. A bigger pump operating well below maximum capacity is quieter than a smaller one operating near maximum capacity. Properly decoupled, it can be pretty quiet (about 31-33dBA for a Laing D5 Vario, which draws about 30W). A good overview of pump noise can be found at Martin’s Liquid Lab: Pump Noise Testing Round 1.

The one thing that worries me about liquid cooling is that the tubing will get brittle with time and develop leaks. I have no idea how long this will take. It may not be an issue for guys who take their machines apart every few months, but once I get a machine set up the way I like it, I prefer not to open the box for years at a time.

Anyway, the idea is to leave plenty of headroom for power and cooling, so I can hopefully run at full load without spinning up the fans very far above idle. It seems to me, that is the key to quiet operation.

And, after taking all that the into account, our estimates are still somewhere around 200W apart ... so now what?

There's no way those power numbers are correct witthout even looking at them. Intel, the people who make the chip rate it at 130W and you'd think they'd know . And would any chip idle at 87W?

Those values are all for the system, not the CPU itself. This means it already includes a motherboard, RAM, SSD, Graphics card and losses from the PSU itself. I think this explains some more earlier in that review:
http://www.tomshardware.com/reviews/cor ... 071-5.html

So, that's 167W for the whole system. OK, the graphics card is not under load and no HDD and DVD drive are included but that will strongly agree with the 400-450W realistic max that CA_Steve previously suggested.


Sorry, you're clearly taking the total system power consumptions again as representative as the motherboard itself. The graphics themselves say global power consumption. These motherboard themselves will be using a small fraction of it.

Please don't buy these motherboards anyway! 'High end' gaming motherboards are just marketing gimmicks. Take the top Intel chipset, lob some cheap addons at it, give it horrid colours, gimmicky features and a higher price and people will buy it, just because it's more expensive. Doesn't make it any better in actual use.


Please can you give some link to where you found a 30W hard disk this side of the 1970s. No modern hard disk uses that power. 10W is a fair load number for a modern HDD and SSDs are any a couple of Watts even at load.

A couple of things to think about:
- When measuring CPU power consumption, most reviewers are measuring system power use and not at the component as edh pointed out.
- When measuring power consumption, most reviewers are measuring at the AC wall socket and not the DC power drawn from the PSU. To find out the DC draw, you need to know the PSU's efficiency at that particular load.

So, using the first Tom's Hardware link, the stressed system load (Prime 95 small FFT) is 253W (AC). Tom's used a Cooler Master UCP-1000W 80 Plus Silver rated PSU. didn't see any 1000W reviews out there with efficiency results, but did find one at Anandtech for the 900W version with ~84% @ 20% load. Using the same results, the 253W (AC) turns into 213W (DC).

This system had a GTX 580 in it...which would be idling for the Prime95 test. An idling GTX 580 draws ~30W. 213W-30W = 187W (DC).

Add the GTX 680 (244W), and maybe 20W for your additional components, and you get 477W (DC) as a stressed load...and maybe 425W while gaming*. Again, if you overvolt/overclock it needs to be addressed.

hope that helps.

* I hope this isn't a gaming build...SNB-E adds no value for gaming.

For grins, here's a couple of reviews of the GTX 680 where the test setups were similar to yours:
X-bit labs had 456W (AC) while running an overclocked and overvolted i7 and overclocked GPU under Metro 2033. The PSU used is ~85% efficient at this load, leading to ~390W (DC) load.

Hot Hardware saw 393W (AC) under heavy workload at stock clocks. They fail to say what PSU is used. Best case, it's Platinum and running at max efficiency..call it 92%. This implies DC load of 362W.

Well. That removes the inconsistency I thought I was seeing in the numbers.

Thank you both for giving me the benefit of your experience.

But who would have thought that a processor review would report power consumption for the whole system? I mean, it’s a nice number to know, but when I’m doing a power and cooling design, I want to know idle, active and peak current draws for each component on each rail, and how much heat each component dissipates under various operating conditions.

The more I look, the harder this information is to find. So, what, do they expect you to do – plug together a bunch of arbitrary components, chosen based on misleading reviews and incomplete specifications, and pray they all play nice together?

Well, maybe not quite. Here’s an interesting overview, though it dates from 2009: Calculating Power Consumption Of The Entire System.

I stand corrected. Going back to the spec sheets, I find 11W for the Western Digital Caviar Black, 3W for the OCZ Vertex 3, and no number for the LG BH14NS40, so I’ll take it to be 9W, the midpoint of Tom’s range of 5-12W.

Putting all the updated numbers together, I get

mobo: 50 W x 1 = 50 W
cpu: 130 W x 1 = 130 W
gpu: 244 W x 1 = 244 W
ram: 0.5 W x 16 = 8 W
disk: 11 W x 1 = 11 W
ssd: 3 W x 1 = 3 W
dvd: 9 W x 1 = 9 W
fan: 3 W x 6 = 18 W
pump: 30 W x 1 = 30 W
overall total: 503 W

So now we are only 26W apart, and that’s close enough for me.

I can’t follow your link to X-bit Labs, because it crashes my browser, but yes, the system reviewed by Hot Hardware looks like it should be within 100W of what I have in mind.

That’s a nice reality check; thanks again.



No, this is not primarily a gaming build; its mostly for mathematical modeling and combinatorial algorithms. The GTX 680 is for its CUDA cores. However, it would be nice to play the odd game, too.

While the 9360X is overkill for today, I want to be able to run the box for years without further hardware upgrades. I like to wait until I can get at least a factor of 10 performance improvement before I move from one system to the next, meaning 6 to 7 years. The only major component I may have to replace in that period is the hard disk.

I may want to open the box at the halfway point to speed up the clocks, though.

And I still have no idea what longevity to expect of a water cooling set-up.

What are your thoughts on these issues?

Unfortunately if you consider how they might accurately measure components it is not possible to measure the CPU itself unless you somehow connect a metre across the right pins of the CPU. They can only really measure the whole system. There are some graphics card reviews where the PCI-E slot power is blanked off and they measure just the 12V so these are accurate but for CPU this is not possible.


If they were really reliable long term then would you expect servers and workstations to use water cooling? Water cooling is almost unheard of in the professional arena so this gives you some idea perhaps.

Phew.

I'm not a fan of water cooling. I know they've taken long strides with reliability of sealed systems...but it's hard to wrap my head around having a consumer product with conductive liquid inside my PC and not worry about it leaking sometime down the road and causing catastrophic damage.

It might be worth mentioning that (for Linux, anyway), having a second video card is convenient for CUDA debugging unless you want to code with X disabled. I guess I don't know about Windows SDK stuff though.

I don't know if you can use integrated graphics for this. I use a cheap second card, myself.

If you have not done much with CUDA programming in the past, it might be worth looking into this. Or if you just plan to use existing CUDA-enabled tools, ignore this

Also:
This
I don't really understand the advantages of water cooling at stock voltages.
edit: I should say this is particularly true with the popular deployment of heat-pipe coolers.

If I were in the business of measuring such things, I would estimate CPU power dissipation from heat outflow. This could done by measuring the rate of heat flow through a heatsink of standardized thermal conductivity to a standardized thermal reservoir, and estimating heat flow through the PC board by measuring the gradient of the temperature of the board around the processor. Power dissipation estimates obtained in this way would be smoothed by the thermal time constant of the processor package, but that's probably just what you want for a power and cooling design.

Also, to get a finer granularity of insight into system power dissipation, you can couple inductively to each cable coming out of the power supply. You would have to start from power-on and integrate changes in current, though, because these are DC currents.
Good point.
Actually, you can get electrically non-conductive coolant. The real problem is that all the most expensive chips in your PC will melt when the coolant drains out of the system. Not to mention the stains caused by coolant seeping into the nice hardwood floor the machine is sitting on.

Am I to conclude, then, that there is no such thing as a quiet, reliable PC with a GTX 680 in it?

Quiet; reliable; GTX 680. Pick any two.

Is that how it goes?

There is no integrated graphics on a 3960X, so this is a good catch. Particularly because, no matter how many times I tell myself to just sit back and use the preconfigured tools, I know I can't resist the urge to program on the bare silicon.
Heat-pipe coolers? That sounds intriguing... do they come in configurations that could conduct the heat away from a CPU and a GPU and dump it to an external heat exchanger?

Well, the CPU and GPU will instantly throttle when they hit specific temp thresholds. So, death is not likely from failed coolant or fan failure in an air cooler....and your BIOS can be set to trigger alarms for both temp and low fan rpm.

Usually, we get a "help me select components for a quiet PC" request. With you, it's 'help me pick out an outrageously high power quiet PSU'...and then we spiraled out to some of the other components. Maybe we can start over.

Yes, you can build a pretty quiet PC with a GTX 680. Asus Direct CU II and MSI Twin Frozr IV solutions are pretty good. You could get a top shelf CPU cooler and it'll be pretty quiet. Get a mobo with decent fan controls or one that'll work with speedfan. Next, get a case with great airflow and quiet fans you are set. If you like, we can suggest some builds.

heat-pipe coolers refers to any air cooler that has a base with heatpipes attached. The heatpipes route through an attached large radiator that has a fan attached. See any CPU cooler review on this site for a reference.

This is an interesting idea. In most of my (limited) PCB design experience, I have striven to use the ground plane as a heatsink for thermally significant components. I assume they do this on motherboards? Maybe not though, since a dedicated cooler is guaranteed. Are most motherboards 4 layer? Is there a ground plane? Or just traces?
Integrated NB/graphics could complicate things, too, but I guess you would get real figures rather than theoretical ones, which is a good thing?

I just meant the little copper tubes which are now on most after market coolers and are even in some primary coolers. I'm still slightly shocked by how much they have changed the air cooling game. They work by having a fluid inside them boil at the heat source and condense on the rest of the pipe. Vapor heat transfer ftw

edit:grammar fail

Sorry about the delay in getting back to you.
Well, I guess that’s true, but we are still left with stains on the floor. Be that as it may, I do prefer air cooling, for both cost and reliability; the only reason I was considering water cooling was because it seemed to be the only way to make a really powerful machine reasonably quiet.
Yeah, I sort of highjacked my own thread, didn’t I? The thing is, all these issues ramify into each other. Initially, I thought I had it pretty much figured out: I would water cool the CPU and GPU, and I would run a PSU overrated by a fgactor of 2, resulting in a powerful but still quiet machine. Now that you’ve set me straight on power requirements, the changes in my thinking are rippling through all the major components.
So far, your suggestions have been right on the money, so I would be very pleased to hear more of them!

I found a couple of reviews of the Asus GTX 680 DirectCU II. Frankly, the one here at silentpcreview is much better than the one at guru3d. The people at guru3d used a test system that put out 37dbA at idle, and gave a figure of 40dbA under load. In other words, only 50% of the noise being measured was coming from the GPU. By comparison, silentpcreview used a test system that put out 15dbA at idle and 28dbA under load, meaning that 95% of the measured noise came from the GPU, which is ten times more accurate than guru3d’s measurement. The end results are dramatically different: according to guru3d, the card puts out 37dbA under load, while silentpcreview measures it at just under 28dbA. The latter is a number I can certainly live with.

The MSI card with Twin Frozr IV cooling is the GTX 680 Lightning. Sadly, I found no review of this card on silentpcreview, though I did find one at guru3d. The same caveat applies, with the test system putting out 36dbA at idle and 38dbA under load. This works out to just under 34dbA noise output from the card, but the measurement is so inaccurate that it isn’t trustworthy.

The only conclusion I can come to is that the MSI GTX 680 Lightning may be slightly quieter than the Asus GTX 680 DirectCU II. The MSI card is 1.9" wide, while the Asus card is 2.3" wide, meaning both require 3 expansion slots, but the MSI card fits much more comfortably, leaving more room for air flow around the card. It’s difficult to say which card to choose. Around the web, no one seems able to argue decisively for one or the other – people usually recommend the card made by the manufacturer they like best.

Asus also makes a 4GB version of the GTX 680 DirectCU II (with a slightly slower clock frequency), which – oddly enough – is only 1.7" wide. I don’t understand the difference in width between the 2GB and the 4GB Asus cards. I lean towards this card because bigger memory means less PCIe bus traffic, which results in a performance boost for large numerical applications that much more than offsets the slight loss due to the clock frequency difference, and because it leaves the most room for airflow.

By the way, here’s 4GB a fanless GTX 680 (!) by Colorful, called the iGame Kudan GTX 680. A few aditional details may be found here and here. Does anybody know anything about this company?

Anyway, I’m interested to know what other ideas you have for a blindingly fast, whisper quiet machine. I guess I should start another thread for that. Would the System Advice / Troubleshooting forum be a good place for it?

So... does anybody make a heat-pipe solution that conducts the heat right out of the box to an external radiator?
Actually, modern high-end motherboards typically have either 6 or 8 layers, half of which are power or ground planes, which will conduct a non-negligible amount of heat away from the CPU. This is why I suggested measuring the temperature gradient on the board around the processor. Knowing that, and estimating the thermal conductivity from the number and weight of power/ground planes, one can make an acceptable estimate of the heat flowing out via the board. This number will be a fair bit smaller than the figure measured via the heat sink, but not so small that it can be neglected entirely. And then there is the heat radiated from the other side of the PC board... this would have to be estimated from the temperature of the boad at that position and the rate of air flow across it.
Sure. But I would still like to know if anybody makes a heat-pipe solution that conducts the heat right out of the box to an external radiator!

For CPUs:
The Phenom 2 factory coolers had heat pipes. It may be that the higher-end (I7? Maybe I5?) Intel coolers have them. But I would personally go for an OEM part and a Cooler Master 212 EVO (or whatever they are calling it now). That part is ~$30, and is leagues better than any boxed CPU cooler. And OEM parts tend to be $10-$20 cheaper, so $30 is pretty good. (These numbers all assume US pricing. I don't see a location in your profile.)
Newegg Cooler Master 212 EVO:
http://www.newegg.com/Product/Product.a ... 6835103099

For GPUs:
As CA Steve was saying, the ASUS Direct CU cards have heat pipes. (The "Direct CU" moniker refers to direct contact between the heat spreader and the heat pipes.) Most good coolers do these days, actually. But I would definitely look for a graphics card with large fan(s). A glance at the 680 on Newegg, and everyone 680 I look at has them. This does not make the coolers equal, though since heat pipe, fin, and fan design are still an issue. The Gigabyte cards look cute (with their 3 fans), but god knows what RPM they run at. I guess some reviews might discuss it. Asus is a good default, though, if you have $30 to spare. Their customer service has been good in my experience (for RMA), and their video cards are generally more quiet than most alternatives these days.

If you'd like more input from other members, then by all means create a new thread in the System Advice forum with a relevant subject line like "Building a quiet/powerful PC for computational analysis" or something. Link back to this thread for background.

Noise testing is tricky. You have to realize that nearly all reviewers are operating out of their homes / apartments. It's not like they have an underground lair. Second, very few take the time/ trouble to convert a room to an anechoic chamber as Mike has done or purchased a very low noise microphone. Thus, the 10-11dBA noise floor for SPCR's reviews. That said, there are a couple of places you can go to when there isn't an SPCR review. HT4Uin Germany does a great job with noise testing as well as showing temps for the GPU and VRM circuitry...and while Guru3D has a high noise floor, you can get a feel for the relative loudness of one card vs another. Techpowerup is another decent site. By the way, it's important to understand how each review site sets up their testing as procedures vary greatly. In Guru3D's case, their calibrated mic is 75cm away from the PC. At SPCR it's 100cm. Inverse square law applies.

As far as heat pipes leading to radiators outside the case....there are systems where the heatpipes [url=http://www.silentpcreview.com/HDPLEX_H5.TODDuse the case itself [/url]as the passive radiator. But, this is limited to much lower power systems.

As for that Colorful card, good luck. While the card itself doesn't have a fan, I'll guarantee it is relying upon a certain amount of airflow in order to work.

I think I will do just that.
I noted that. But 75cm vs. 100cm only accounts for about 2.5 db, which doesn’t get anywhere near accounting for the difference between the two measurements. So you’re right, the best you can get from guru3d is a qualitative feel for noise levels.
Of course it is. But there would be fans moving air through the case in any event. The GPU fans are generally the loudest ones in the system. If they could be eliminated, and their job could be done instead by big, slow-spinning case fans -- without throttling down the GPU -- it would be a real bonus. Only time will tell, I guess.

Colorful: 200W is still 200W. Hard to cool that w/o a fan attached. Heck, it's hard to do it at 100W.

I believe for SPL the formula would be 20log10[(75^2)/(100^2)] or roughly 6dB. Yes, still no where near the same.

@Colorful

Before we have it reviewed here, we can't say for sure, but i'll highly doubt you'll be able to conduct 200W of heat away from this card without any kind of case fans. In addition, look at the size of that monster... you'll probably need a truck garage for a case.

The highest powered card commonly available with passive cooling was the 8800GT fitted with an Accellero S1 rev 2. That card was 126W. The more modern S1 Plus might allow a little more headroom and there is always the ability to push to higher temperatures and maybe then 150W would be workable but not 200W.

Actually, sound pressure falls off as 1/r. It's sound intensity which falls off as 1/r^2. So the correct formula is 20 log(75/100) = 2.5db. Think of it this way: in the analogy between acoustics and electricity, sound pressure corresponds to voltage, flow rate (aka volume velocity) corresponds to current, and their product is sound intensity, which corresponds to power.
Of course case fans will be required. But in a reasonable case, we are talking about multiple large, slow-spinning fans, which can move a lot of air pretty quietly, as compared to the smaller, much louder fans on a typical GPU. You’re right about the size, though. It looks like a 4- or 5-slot card.

All told, I think CA_Steve's suggestion, the Asus GTX 680 DirectCU II, at 28dbA max, is going to be hard to beat for an optimal trade-off between quietness, reliability, and computation. But even that is a 3-slot card. So you need a big motherboard in any case -- especially if you want to have room for two of them. I don't mind going to a full tower case, though a mid tower would be preferable.
Well, anything can be cooled with a large enough radiating area. See Pappnaas' remark about the size of that monster. Given that, Colorful's claim can't be dissmissed out of hand, but you can rest assured that I won't by buying one without first seeing some credible reviews.

Well, "almost" should be the keyword here. I'm in computer graphics and animation, and there are a number of workstation builders that do use sealed water cooling with factory overclocked systems (among the most famous in that market, BOXX comes to mind). I've been doing it myself for at least three years without a single complaint from my customers. Sure, water cooling used to be a geeky thing. Today, it's probably the best choice for system integrators and small builders, since huge air coolers, that will cool the same, are really risky for transportation (much too heavy, flimsy fan clips, etc.).

But I came here looking for a system running four graphics cards for GPGPU use, so I'm also interested in the PSU numbers.

Will a 680 or a TITAN take as much power when running CUDA apps (like a 3D render) as they do when gaming? What's the peak wattage one can expect from these cards in a GPGPU environment? My usual 4.0GHz 3930K overclocked no-graphics system (well, I use an old passive 8500GT for the tests, and for the final build a Quadro 2000 or FirePro equivalent that take around 50W to 75W) draws around 350W at Prime95. If each 680/TITAN drawed 200W each, I would definitely need a 1000+ PSU for a compute workstation.

http://www.guru3d.com/articles_pages/ge ... iew,9.html

If assuming GPGPU load is comparable to furmark regarding power consumption (single non OC card measured at around 250W), 1500W PSU doesn't seem to far off any more.

The fact these systems are factory overclocked already puts it outside of mainstream servers and workstations which are the overwhelming majority of professional systems. Serious professional computing does not overclock and does not use water cooling. Perhaps in some niche areas such as you describe it is acceptable to do so but in terms of long term reliability and liability it is not acceptable and for these reason the major manufacturers or any of the Unix workstation manufacturers would never do so. I stand by my claim.


The GTX Titan has a TDP of 250W and I really wouldn't trust Guru3D's numbers on this, it's typical gaming review site overcompensating. I suggested a method for doing this a while ago, start with one graphics card and a donor PSU and take some measurements, then add in another GPU, then do some maths and work out how big a PSU you need.

I really don't see the worth of the GTX Titan at this point as we are only a matter of months before the next generation cards which will be better and much cheaper.